Multiclass Vulnerability Dataset (MVD) is a database for research on multiclass vulnerability detection with deep learning. The dataset is based on the NIST Software Assurance Reference Dataset (SARD) and National Vulnerability Database (NVD). Up to now, it has possessed 181641 code gadgets, covering 40 types of vulnerabilities. Each code gadget in MVD is composed of multiple program statements, which have direct or indirect data-dependence and control-dependence relationships with the library/API function calls. In total, the code gadgets in MVD are extracted from 33409 testcases of SARD and NVD, 138522 code gadgets of which are non-vulnerable and 43119 are vulnerable.
In this repository, the compressed file mvd.txt.zip stores 181641 code gadgets and their corresponding labels. The file named label2CWE.txt records the mapping relationship between each label and the corresponding vulnerability. The folder source files contains 33,409 source files for extracting code gadgets.